Version 9 (modified by bgamari, 3 years ago) (diff)


Ben's notes on compiler performance

Callstack solver and local lets

3ec8288a18d57fb856e257905897daae237a1d5d (Rework the Implicit CallStack solver to handle local lets) decreased compiler allocations on some tests. Later 7407a66d5bd29aa011f5a4228c6e2b2f7f8ad3f8 (Don't infer CallStacks) undid many of these gains. TODO Why? See, for instance, anna/Dependency, where the former decreases allocations by 38%, and the latter increases them by 25%, with a corresponding ~5% increase in compile time.


See #11196. Unfortunately, the commit of nokinds merge (6746549772c5cc0ac66c0fce562f297f4d4b80a2) had a few issues rendering it untestable. This means that the regressions are attributed to an unrelated commit (b5d5d83122c93c2a25839127edfd6b2df7ed6928, Revert .gitmodules changes from 6746549772c5).

Along with introducing runtime-rep polymorphism, the nokinds merge eliminated the FunTy constructor in favor of a (likely) larger representation. This was undone by SPJ in 77bb09270c70455bbd547470c4e995707d19f37d. This commit was sadly untestable but allocations didn't change terribly much between the two surrounding commits (e368f3265b80aeb337fbac3f6a70ee54ab14edfd and e33ca0e54f3c20a8b233a3f7b38e4968a4955300).

Adding instances

There are a few examples where adding instances to base modules increases compiler allocations, and often compile time, remarkably.

I suspect these will both be addressed by ed480981/ghc, which fixes a bug (introduced by 4c834fdddf4d44d12039da4d6a2c63a660975b95 and perhaps worsened by later commits) in instance resolution which makes interface loading significantly more eager than necessary.

Tuple constraint size

Rather significant regression due to dd3080fe0263082f65bf2570f49189c277b12e28. Ended up being due to a redundant lookup in the original name cache. See #12357.


See #12506.